Model reference · open weights
MiniCPM-OCR-1.0-DSpark is an open-weight language model from openbmb. MiniCPM-OCR-1.0-DSpark (BF16) weighs 595 MB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the openbmb/MiniCPM-OCR-1.0-DSpark-gguf model card, 2026-10-28. The estimate below is for another build of the family.
What it is
| Released by | openbmb |
|---|---|
| Released | 2026-10-26 |
| VRAM | 595 MB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB | 122 | 30 | all 256K | 11.6 GB |
| RTX 4060 Ti 16 GB | 167 | 41 | all 256K | 15.4 GB |
| RTX 3090 24 GB | 263 | 65 | all 256K | 23.4 GB |
| RTX 4090 24 GB | 262 | 65 | all 256K | 23.4 GB |
| RTX 5090 32 GB | 353 | 88 | all 256K | 31.0 GB |
| L40S 48 GB | 508 | 127 | all 256K | 44.0 GB |
| A100 80 GB | 916 | 229 | all 256K | 78.2 GB |
| H100 80 GB | 855 | 213 | all 256K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 1000+ | 260 | all 256K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 1000+ | 300 | all 256K | 107 GB |
| H200 141 GB | 1000+ | 392 | all 256K | 138 GB |
| B200 180 GB | 1000+ | 505 | all 256K | 176 GB |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 1.4 GB | 1.7 GB |
| 5 | 1.7 GB | 3.0 GB |
| 8 | 2.0 GB | 4.0 GB |
| 16 | 2.7 GB | 6.7 GB |
| 32 | 4.0 GB | 12.1 GB |
| 64 | 6.7 GB | 22.8 GB |
One card, with vLLM's small-card settings.
Builds
| Build | Params | Precision | Weights | Smallest setup |
|---|---|---|---|---|
| MiniCPM-OCR-1.0-DSpark-gguf ↗ 11 builds: Q4_0 209 MB … F16 606 MB |
GGUF | 211 MB | RTX 3060 12 GB | |
| MiniCPM-OCR-1.0-DSpark (above) ↗ | 297M | BF16 | 595 MB | RTX 3060 12 GB |
From the model card
GGUF build of the DSpark draft head for MiniCPM-OCR 1.0. A draft model, not a standalone OCR model — pair it with the matching main checkpoint.
| File | Size |
|---|---|
dspark-MiniCPM-OCR-1_0-F16.gguf | 0.606 GB |
dspark-MiniCPM-OCR-1_0-Q8_0.gguf | 0.327 GB |
dspark-MiniCPM-OCR-1_0-Q6_K.gguf | 0.255 GB |
dspark-MiniCPM-OCR-1_0-Q5_K_M.gguf | 0.233 GB |
dspark-MiniCPM-OCR-1_0-Q5_K_S.gguf | 0.232 GB |
dspark-MiniCPM-OCR-1_0-Q5_1.gguf | 0.243 GB |
dspark-MiniCPM-OCR-1_0-Q5_0.gguf | 0.232 GB |
dspark-MiniCPM-OCR-1_0-Q4_K_M.gguf | 0.211 GB |
dspark-MiniCPM-OCR-1_0-Q4_K_S.gguf | 0.209 GB |
dspark-MiniCPM-OCR-1_0-Q4_1.gguf | 0.220 GB |
dspark-MiniCPM-OCR-1_0-Q4_0.gguf | 0.209 GB |
Use the main model's projector (mmproj-model-f16-4x.gguf from
MiniCPM-OCR-1.0-gguf); this repository contains only the draft.
llama-cli \
-m MiniCPM-OCR-1_0-Q8_0.gguf \
--mmproj mmproj-model-f16-4x.gguf \
-md dspark-MiniCPM-OCR-1_0-Q8_0.gguf \
--spec-type draft-dspark \
--image page.png -p "Text Recognition:" \
--jinja -ngl 99 -ngld 99 -fa on
-md is the draft model and --spec-type draft-dspark enables the draft; -ngld offloads draft
layers to GPU.
llama-mtmd-clidoes not accept a draft model. Usellama-cliorllama-serverwhen enabling DSpark with multimodal input.
Each tier carries the absolute target token IDs in its d2t tensor (the safetensors export
stores them delta-encoded).
| Format | Hugging Face | ModelScope |
|---|---|---|
| BF16 (original) | 🤗 | ModelScope |
| GGUF (llama.cpp) | 🤗 | ModelScope |
| AWQ W4A16 | 🤗 | ModelScope |
| GPTQ W4A16 | 🤗 | ModelScope |
| DSpark draft (BF16) | 🤗 | ModelScope |
| DSpark draft (GGUF) (this repo) | 🤗 | ModelScope |
Installation, model downloads, GPU memory, input formats, and unexpected outputs: FAQ.
Model weights, code, and documentation are released under the Apache License 2.0.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works